OpenAI launches GPT-5.6 flagship in three versions: Luna, Terra, and Sol

On July 17, 2026, OpenAI made its flagship GPT-5.6 model family generally available in three sizes at once: Luna, Terra, and Sol. According to OpenAI's own benchmark, the top-tier Sol beat Anthropic's Claude Fable 5 by 13.1 points on Agents' Last Exam, a test of long-running agentic tasks.

How much do the new models cost?

Pricing is set per 1 million tokens, with separate rates for input and output, and the three GPT-5.6 sizes differ noticeably in cost.

• Luna (smallest model): $1 per 1M input tokens / $6 per 1M output tokens
• Terra (mid-tier model): $2.50 per 1M input tokens / $15 per 1M output tokens
• Sol (top-tier model): $5 per 1M input tokens / $30 per 1M output tokens
• For comparison, the Claude Opus line costs $5/$25 per 1M tokens, and Claude Fable 5 costs $10/$50
• All three GPT-5.6 models have a knowledge cutoff of February 16, 2026, a 1 million-token context window, and a 128,000-token output limit

How far ahead of the competition is GPT-5.6?

OpenAI leans on Agents' Last Exam, a benchmark of long-running agentic workflows across 55 professional domains. The company says GPT-5.6 Sol scored 53.6 points on the test, a new high, beating Claude Fable 5 (in adaptive reasoning mode) by 13.1 points. Even at medium reasoning, Sol leads Fable 5 by 11.4 points at roughly a quarter of the cost. OpenAI claims the smaller models, Terra and Luna, also beat Fable 5 while costing about 16 times less to run.

On another popular benchmark, though, the picture flips. On SWE-Bench Pro, which tests models on real-world software engineering tasks, Claude Fable 5 scored 80%, while GPT-5.6 Sol managed only 64.6%. That may explain why, on the eve of the launch, OpenAI published a separate article saying it considers about 30% of SWE-bench Pro tasks to be broken and urging other model developers to scrutinize results on the benchmark more carefully.

What do early tests show?

A blogger who had early access to GPT-5.6 Sol describes the model as "definitely very competent," but notes that on the hard coding tasks he works on, it has not yet struck him as better than Anthropic's Claude Fable 5. In his view, the most interesting details are in OpenAI's official GPT-5.6 usage guide, which describes a number of new API capabilities that third-party tool developers have yet to explore.

What does it mean?

The three-tier GPT-5.6 launch continues the race between OpenAI and Anthropic for the lead in agentic tasks, where a model carries out long chains of actions on its own rather than just answering a one-off prompt. The benchmarks, however, paint a contradictory picture: OpenAI wins on its own agentic-workflow test but loses on SWE-Bench Pro, a benchmark whose task quality OpenAI itself has publicly questioned for a third of its tasks. The takeaway for developers: test models on your own workloads instead of relying on a single averaged leaderboard.

Frequently asked questions

How much does GPT-5.6 Sol cost?

The top-tier GPT-5.6 Sol costs $5 per 1 million input tokens and $30 per 1 million output tokens. That makes it pricier than Luna ($1/$6) and Terra ($2.50/$15), but cheaper than Claude Fable 5 ($10/$50).

How does GPT-5.6 compare with

Claude Fable 5 in capabilities?

On the Agents' Last Exam benchmark, GPT-5.6 Sol leads Fable 5 by 13.1 points (53.6 versus Fable 5's score), but on SWE-Bench Pro, Fable 5 scores 80% to Sol's 64.6%. Which model comes out ahead depends on the type of task.

What are the context window and limits of the GPT-5.6 models?

All three models, Luna, Terra, and Sol, have a 1 million-token context window, a 128,000-token output limit, and a knowledge cutoff of February 16, 2026.